Table of Contents

• 1  COVSMA's first tool: COVSCO
  • 1.1  Data Exploration
    • 1.1.1  The data
      • 1.1.1.1  The features: 22 variables from which we will predict our target
      • 1.1.1.2  The Target: New hospitalizations due to severe COVID19 cases
    • 1.1.2  New hospitalizations means over all departements in function of 7davg pollutants concentrations differencials
    • 1.1.3  New hospitalizations means over all departements in function of 1MMax Pollutants concentrations differencials
    • 1.1.4  Ozone (O3) and the number of severe COVID19 cases leading to hospitalization
      • 1.1.4.1  Departement 75: Paris region Ile de France
      • 1.1.4.2  Departement 83: Var region PACA
    • 1.1.5  Nitrogen dioxide (NO2) and the number of severe COVID19 cases leading to hospitalization
      • 1.1.5.1  Departement 75: Paris region Ile de France
      • 1.1.5.2  Departement 83: Var region PACA
    • 1.1.6  PM2.5 and the number of severe COVID19 cases leading to hospitalization
      • 1.1.6.1  Departement 75: Paris region Ile de France
      • 1.1.6.2  Departement 83: Var region PACA
    • 1.1.7  CO and the number of severe COVID19 cases leading to hospitalization
      • 1.1.7.1  Departement 75: Paris region Ile de France
    • 1.1.8  The most polluted departements of France
      • 1.1.8.1  The Ozone O3 Pollutant
      • 1.1.8.2  The PM2.5 Pollutant
      • 1.1.8.3  The NO2 Pollutant
      • 1.1.8.4  The CO Pollutant
      • 1.1.8.5  The PM10 Pollutant
    • 1.1.9  The facebook mobility index
      • 1.1.9.1  The facebook mobility index VS Ozone O3
      • 1.1.9.2  The facebook mobility index VS Carbon Monoxyde CO
      • 1.1.9.3  The facebook mobility index VS Nitrogen Dioxide NO2
      • 1.1.9.4  The facebook mobility index VS PM2.5
      • 1.1.9.5  The facebook mobility index VS PM10
  • 1.2  Training the model - Gradient Boosting for regression
    • 1.2.1  Hold-out and Cross Validation (MSE/MAE)
    • 1.2.2  Feature importance Report
    • 1.2.3  Exporting the model to a joblib file
    • 1.2.4  Running T-POT Auto ML optimization algorithm
    • 1.2.5  Recurrent Neural Network
  • 1.3  Conclusion

COVSMA's first tool: COVSCO

COVSMA stands for Copernicus Satellites Versus Maladies: The current sanitary crisis implied the necessity to develop an online tool to monitor pollution levels, display alerts that will imply that governments automatically take measures: days without cars and trucks that are not 100% Electrical, monitor the live impact of the measures taken, and forecast COVID19 risk for up to 4 days in the future, COVID19 risk being defined as the predicted numbers of new hospitalisations due to severe COVID19 cases for all states/departements. We have baptised this tool with the analog name of COVSCO (Copernicus Satellites Versus COVID19). We start with France and its 96 departements. A follow up will be to apply the same methodology to severe respiratory diseases and to expand the model and databases to a worldwide scale.

Data Exploration

The data

              nom  numero       time  hospi  reanim  newhospi  newreanim  \
34            Ain     1.0 2020-05-14  137.0     8.0       4.0        0.0   
35            Ain     1.0 2020-05-15  135.0     7.0       4.0        0.0   
36            Ain     1.0 2020-05-16  134.0     6.0       1.0        0.0   
37            Ain     1.0 2020-05-17  133.0     6.0       1.0        0.0   
38            Ain     1.0 2020-05-18  132.0     6.0       1.0        0.0   
...           ...     ...        ...    ...     ...       ...        ...   
34171  Val-d'Oise    95.0 2021-03-27  659.0    79.0      43.0        7.0   
34172  Val-d'Oise    95.0 2021-03-28  667.0    81.0      38.0        7.0   
34173  Val-d'Oise    95.0 2021-03-29  680.0    76.0      44.0        4.0   
34174  Val-d'Oise    95.0 2021-03-30  688.0    75.0      88.0        5.0   
34175  Val-d'Oise    95.0 2021-03-31  698.0    77.0      83.0        7.0   

        deces  gueris  dep_num  ...  depnum_y  Smokers Nb_susp_501Y_V1  \
34       88.0   318.0      1.0  ...         1    0.262               0   
35       89.0   323.0      1.0  ...         1    0.262               0   
36       90.0   325.0      1.0  ...         1    0.262               0   
37       90.0   326.0      1.0  ...         1    0.262               0   
38       90.0   331.0      1.0  ...         1    0.262               0   
...       ...     ...      ...  ...       ...      ...             ...   
34171  1603.0  6857.0     95.0  ...        95    0.213            9561   
34172  1606.0  6882.0     95.0  ...        95    0.213            9557   
34173  1615.0  6902.0     95.0  ...        95    0.213            9496   
34174  1632.0  6964.0     95.0  ...        95    0.213            8643   
34175  1639.0  7024.0     95.0  ...        95    0.213            7915   

      Nb_susp_501Y_V2_3  minority  pauvrete  rsa  ouvriers  \
34                    0   54821.0      10.7  2.3     17.74   
35                    0   54821.0      10.7  2.3     17.74   
36                    0   54821.0      10.7  2.3     17.74   
37                    0   54821.0      10.7  2.3     17.74   
38                    0   54821.0      10.7  2.3     17.74   
...                 ...       ...       ...  ...       ...   
34171               432  161947.0      16.8  5.8     17.63   
34172               428  161947.0      16.8  5.8     17.63   
34173               396  161947.0      16.8  5.8     17.63   
34174               364  161947.0      16.8  5.8     17.63   
34175               332  161947.0      16.8  5.8     17.63   

       totalcovidcasescumulated  prevdaytotalcovidcasescumulated  
34                         36.0                               18  
35                         46.0                               36  
36                         46.0                               46  
37                         46.0                               46  
38                         68.0                               46  
...                         ...                              ...  
34171                  225860.0                           223604  
34172                  226188.0                           225860  
34173                  230717.0                           226188  
34174                  234429.0                           230717  
34175                  237793.0                           234429  

[30912 rows x 93 columns]

	idx	pm25	no2	o3	pm10	co	pm257davg	no27davg	o37davg	co7davg	...	normno27davg	normo37davg	normpm107davg	normco7davg	normpm251Mavg	normno21Mavg	normo31Mavg	normpm101Mavg	normco1Mavg	newhospi
34	631877.0	6.403166	3.939003	43.262828	7.967417	177.622754	7.257504	3.476745	72.813923	161.817954	...	0.046602	0.586532	0.129071	0.234867	0.108579	0.048632	0.559299	0.094840	0.218359	4.0
35	631877.0	10.041256	4.039703	45.365958	12.985050	177.204810	7.283296	3.523846	71.713745	162.361518	...	0.047305	0.577448	0.127923	0.236187	0.119927	0.051150	0.496692	0.102941	0.226082	4.0
36	631877.0	8.650893	2.993409	64.447998	10.213892	173.986833	7.222348	3.502823	70.928693	162.525890	...	0.046991	0.570966	0.125429	0.236587	0.129652	0.051861	0.455272	0.105411	0.234060	1.0
37	631877.0	7.924968	2.470320	81.736362	11.228378	163.052671	7.159819	3.487527	70.520307	162.329540	...	0.046763	0.567594	0.123235	0.236110	0.127766	0.051226	0.458184	0.101608	0.239713	1.0
38	631877.0	8.803713	2.883282	79.918855	11.338718	186.330957	7.143561	3.477693	69.999498	162.848931	...	0.046616	0.563293	0.120713	0.237371	0.142927	0.049201	0.477197	0.113024	0.255846	1.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
34171	1215390.0	6.100140	6.627001	70.229467	11.613934	185.482428	14.580851	14.493739	56.915039	207.123258	...	0.211142	0.455257	0.281847	0.344925	0.302061	0.237179	0.471712	0.271828	0.352307	43.0
34172	1215390.0	8.425896	9.099569	68.209629	13.743489	193.139682	14.195306	14.223872	57.801110	206.960739	...	0.207112	0.462573	0.269762	0.344530	0.289930	0.240049	0.480526	0.265162	0.358714	38.0
34173	1215390.0	13.033086	29.526526	54.441418	18.517216	237.388093	14.249303	14.775230	57.704602	207.544318	...	0.215346	0.461776	0.270202	0.345948	0.295797	0.280570	0.463981	0.266083	0.371198	44.0
34174	1215390.0	21.792762	51.437356	40.131381	30.011262	307.382279	14.525749	16.057324	57.291931	210.476992	...	0.234495	0.458369	0.274637	0.353072	0.286940	0.319473	0.461757	0.261254	0.379313	88.0
34175	1215390.0	21.862512	49.729555	42.796275	32.744934	290.497534	14.654314	17.360417	57.019812	213.319222	...	0.253957	0.456122	0.278918	0.359977	0.278900	0.366677	0.452311	0.260245	0.393208	83.0

30912 rows × 56 columns

The features: 22 variables from which we will predict our target

Index(['idx', 'pm25', 'pm257davg', 'normpm25', 'hospiprevday',
       'covidpostestprevday', 'prevdaytotalcovidcasescumulated',
       'all_day_bing_tiles_visited_relative_change',
       'all_day_ratio_single_tile_users', 'vac1nb', 'vac2nb',
       'Insuffisance respiratoire chronique grave (ALD14)',
       'Insuffisance cardiaque grave, troubles du rythme graves, cardiopathies valvulaires graves, cardiopathies congénitales graves (ALD5)',
       'Smokers', 'minority', 'Nb_susp_501Y_V1', 'Nb_susp_501Y_V2_3',
       '1MMaxpm25', 'pm251Mavg', 'pauvrete', 'rsa', 'ouvriers'],
      dtype='object')
22

The Target: New hospitalizations due to severe COVID19 cases

The daily number of new hospitalizations due to severe COVID19 cases for every French departement is what we will predict.

New hospitalizations means over all departements in function of 7davg pollutants concentrations differencials

Text(0, 0.5, 'newhospimean')

New hospitalizations means over all departements in function of 1MMax Pollutants concentrations differencials

Text(0, 0.5, 'newhospimean')

Ozone (O3) and the number of severe COVID19 cases leading to hospitalization

Departement 75: Paris region Ile de France

<AxesSubplot:>

Departement 83: Var region PACA

<AxesSubplot:>

Nitrogen dioxide (NO2) and the number of severe COVID19 cases leading to hospitalization

Departement 75: Paris region Ile de France

<AxesSubplot:>

Departement 83: Var region PACA

<AxesSubplot:>

PM2.5 and the number of severe COVID19 cases leading to hospitalization

Departement 75: Paris region Ile de France

<AxesSubplot:>

Departement 83: Var region PACA

<AxesSubplot:>

CO and the number of severe COVID19 cases leading to hospitalization

Departement 75: Paris region Ile de France

<AxesSubplot:>

The most polluted departements of France

The Ozone O3 Pollutant

[95.]
                Département  Numéro Date of pollution peak     1MMaxo3  \
0                Val-d'Oise    95.0             2020-08-09  122.889769   
1   Alpes-de-Haute-Provence     4.0             2020-08-08  119.833327   
2                 Haut-Rhin    68.0             2020-07-31  116.979563   
3                   Moselle    57.0             2020-08-10  115.873423   
4                       Ain     1.0             2020-09-18  114.506434   
..                      ...     ...                    ...         ...   
90     Pyrénées-Atlantiques    64.0             2020-08-06   94.093549   
91                     Gard    30.0             2020-05-20   93.974202   
92                     Cher    18.0             2020-07-13   93.921768   
93          Hautes-Pyrénées    65.0             2020-05-30   93.086194   
94                   Ariège     9.0             2020-05-29   90.297432   

    totalcovidcasescumulated  Population Index  
0                     237793         1215390.0  
1                      20858          161799.0  
2                      79025          762607.0  
3                     141705         1044486.0  
4                     109223          631877.0  
..                       ...               ...  
90                     62382          670032.0  
91                    107351          738189.0  
92                     32904          308992.0  
93                     25307          228582.0  
94                     14118          152499.0  

[95 rows x 6 columns]

The PM2.5 Pollutant

[59.]
                Département  Numéro Date of pollution peak  1MMaxpm25  \
0                      Nord    59.0             2020-11-27  39.932960   
1                 Haut-Rhin    68.0             2021-02-24  37.243984   
2               Deux-Sèvres    79.0             2021-03-09  36.380767   
3                     Paris    75.0             2021-01-02  35.418335   
4                    Vienne    86.0             2021-03-09  34.895373   
..                      ...     ...                    ...        ...   
90          Alpes-Maritimes     6.0             2021-03-06  20.296409   
91                Côte-d'Or    21.0             2021-02-23  20.115372   
92                   Lozère    48.0             2021-03-04  19.935369   
93  Alpes-de-Haute-Provence     4.0             2021-02-23  19.765781   
94                  Ardèche     7.0             2021-03-04  19.407017   

    totalcovidcasescumulated  Population Index  
0                     477936         2605238.0  
1                      79025          762607.0  
2                      35767          374435.0  
3                     404122         2206488.0  
4                      38030          434887.0  
..                       ...               ...  
90                    219293         1082440.0  
91                     66868          533147.0  
92                     10367           76309.0  
93                     20858          161799.0  
94                     45658          324209.0  

[95 rows x 6 columns]

The NO2 Pollutant

[75.]
            Département  Numéro Date of pollution peak   1MMaxno2  \
0                 Paris    75.0             2021-03-31  67.312539   
1        Hauts-de-Seine    92.0             2021-03-02  64.475306   
2          Val-de-Marne    94.0             2021-01-08  53.384538   
3            Val-d'Oise    95.0             2021-03-02  51.599818   
4              Yvelines    78.0             2020-11-26  43.024886   
..                  ...     ...                    ...        ...   
90               Lozère    48.0             2021-01-07   7.727260   
91         Corse-du-Sud   201.0             2020-12-09   6.285247   
92               Ariège     9.0             2021-01-09   6.254226   
93  Pyrénées-Orientales    66.0             2021-01-09   6.071631   
94          Haute-Corse   202.0             2021-01-09   5.592579   

    totalcovidcasescumulated  Population Index  
0                     404122         2206488.0  
1                     266670         1601569.0  
2                     265897         1372389.0  
3                     237793         1215390.0  
4                     212766         1427291.0  
..                       ...               ...  
90                     10367           76309.0  
91                     11183          152730.0  
92                     14118          152499.0  
93                     45028          471038.0  
94                     13884          174553.0  

[95 rows x 6 columns]

The CO Pollutant

[92.]
            Département  Numéro Date of pollution peak     1MMaxco  \
0        Hauts-de-Seine    92.0             2020-11-26  476.783872   
1              Bas-Rhin    67.0             2020-11-10  442.772829   
2                 Paris    75.0             2020-11-26  442.031472   
3          Val-de-Marne    94.0             2021-01-02  400.354289   
4      Bouches-du-Rhône    13.0             2021-02-24  364.207868   
..                  ...     ...                    ...         ...   
90               Cantal    15.0             2021-01-06  204.699850   
91               Lozère    48.0             2021-01-07  202.482931   
92      Hautes-Pyrénées    65.0             2021-01-10  200.001928   
93               Ariège     9.0             2021-01-11  189.451920   
94  Pyrénées-Orientales    66.0             2021-03-06  182.390438   

    totalcovidcasescumulated  Population Index  
0                     266670         1601569.0  
1                     138675         1116658.0  
2                     404122         2206488.0  
3                     265897         1372389.0  
4                     404460         2016622.0  
..                       ...               ...  
90                     11762          146219.0  
91                     10367           76309.0  
92                     25307          228582.0  
93                     14118          152499.0  
94                     45028          471038.0  

[95 rows x 6 columns]

The PM10 Pollutant

[67.]
     Département  Numéro Date of pollution peak  1MMaxpm10  \
0       Bas-Rhin    67.0             2021-02-25  74.188288   
1      Haut-Rhin    68.0             2021-02-25  71.831104   
2   Corse-du-Sud   201.0             2021-02-06  70.996064   
3         Vosges    88.0             2021-02-25  70.504318   
4    Haute-Saône    70.0             2021-02-25  69.817902   
..           ...     ...                    ...        ...   
90       Mayenne    53.0             2021-03-03  40.262764   
91          Eure    27.0             2021-03-03  39.774691   
92      Calvados    14.0             2021-03-02  38.762629   
93        Sarthe    72.0             2021-03-03  36.942586   
94          Orne    61.0             2021-03-02  36.142811   

    totalcovidcasescumulated  Population Index  
0                     138675         1116658.0  
1                      79025          762607.0  
2                      11183          152730.0  
3                      41055          372016.0  
4                      28170          237706.0  
..                       ...               ...  
90                     29192          307940.0  
91                     65397          601948.0  
92                     61873          693579.0  
93                     57288          568445.0  
94                     29094          286618.0  

[95 rows x 6 columns]

The facebook mobility index

The facebook mobility index VS Ozone O3

<AxesSubplot:>

The facebook mobility index VS Carbon Monoxyde CO

<AxesSubplot:>

The facebook mobility index VS Nitrogen Dioxide NO2

<AxesSubplot:>

The facebook mobility index VS PM2.5

<AxesSubplot:>

The facebook mobility index VS PM10

<AxesSubplot:>

Training the model - Gradient Boosting for regression

Gradient Boosting for regression.

GB builds an additive model in a forward stage-wise fashion; it allows for the optimization of arbitrary differentiable loss functions. In each stage a regression tree is fit on the negative gradient of the given loss function.

Stack of estimators with a final regressor.

Stacked generalization consists in stacking the output of individual estimator and use a regressor to compute the final prediction. Stacking allows to use the strength of each individual estimator by using their output as input of a final estimator.

Note that estimators_ are fitted on the full X while finalestimator is trained using cross-validated predictions of the base estimators using cross_val_predict.

Hold-out and Cross Validation (MSE/MAE)

T-Pot exported current best pipeline
MSE:
48.927229830941876
MAE:
3.542195756660395


{'fit_time': array([21.91892815, 21.4157958 , 28.4483602 , 23.76390553, 26.08099937]), 'score_time': array([0.0273695 , 0.02692819, 0.02848053, 0.05226946, 0.02958846]), 'test_neg_mean_squared_error': array([ -88.03497565,  -67.89317231,  -18.78724034,  -19.55104613,
       -108.93610049]), 'train_neg_mean_squared_error': array([-24.80337586, -28.87515478, -33.52673893, -32.6934117 ,
       -24.36486273]), 'test_neg_mean_absolute_error': array([-4.98415949, -4.4333311 , -2.49463766, -2.54796054, -6.24690705]), 'train_neg_mean_absolute_error': array([-2.5177104 , -2.62877477, -2.93896481, -2.87846415, -2.45653406])}
MSE:
-60.64050698424177
MAE
-4.1413991679799365

Feature importance Report

Scikit Learn - GradientBoostingRegressor:
                                                index  feature_importance
9                                              vac1nb            0.001813
16                                  Nb_susp_501Y_V2_3            0.001920
19                                           pauvrete            0.002501
13                                            Smokers            0.002902
10                                             vac2nb            0.002929
1                                                pm25            0.002964
20                                                rsa            0.003047
3                                            normpm25            0.003323
21                                           ouvriers            0.003725
11  Insuffisance respiratoire chronique grave (ALD14)            0.003786
15                                    Nb_susp_501Y_V1            0.004309
12  Insuffisance cardiaque grave, troubles du ryth...            0.004505
14                                           minority            0.004867
0                                                 idx            0.005109
17                                          1MMaxpm25            0.006621
18                                          pm251Mavg            0.007130
2                                           pm257davg            0.009068
7          all_day_bing_tiles_visited_relative_change            0.011923
8                     all_day_ratio_single_tile_users            0.039006
6                     prevdaytotalcovidcasescumulated            0.122313
5                                 covidpostestprevday            0.274968
4                                        hospiprevday            0.481272

<Figure size 900x600 with 0 Axes>

Exporting the model to a joblib file

Running T-POT Auto ML optimization algorithm

TPOTRegressor

Version 0.11.6.post1 of tpot is outdated. Version 0.11.7 was released Wednesday January 06, 2021.

TPOT closed during evaluation in one generation.
WARNING: TPOT may not provide a good pipeline if TPOT is stopped/interrupted in a early generation.


TPOT closed prematurely. Will use the current best pipeline.
Best pipeline: ExtraTreesRegressor(CombineDFs(input_matrix, input_matrix), bootstrap=False, max_features=0.5, min_samples_leaf=1, min_samples_split=20, n_estimators=100)
-48.56940976377202

Recurrent Neural Network

Conclusion

<seaborn.axisgrid.PairGrid at 0x7fb106271e80>

• Although the virus' contagious characteristic and the mobility index appear to lead in our model's feature importance report, we can notice that unusually high levels in Ozone at ground level are synchronized with the beginning of the epidemy and that the mean of all french departments' new hospitalizations due to severe COVID-19 cases is an increasing function of PM2.5-1-M Maximum and PM10 7-day average differentials.

• We are building a tool, that in addition to predicting with a Mean Absolute error of 4, new departmental hospitalizations due to COVID19 severe cases (COVID Risk) for up to 4 days in the future, and to extending these predictions to hospitalizations due to respiratory diseases when COVID19 is over, ranks French departments by their pollution levels and gives alerts when PM2.5 and PM10 levels are abnormally high for all departments.

• These alerts are translated into recommendations in order that the government takes measures to stop heavy traffic pollution during an amount of time determined by monitoring the live levels in pollutants.

• As the relation that underlies how long the heavy traffic of cars and trucks that aren't 100% electrical has to be stopped for PM2.5 and PM10 to drop to normal levels isn't well known, the satellite data analysis tool we are building and the live graphical visualizations it offers will ensure the best monitoring of how the concentrations in Pollutants (CIPs) go down with the measures the government takes and give confirmations when the CIPs lower to a decent level.